DMMNet: A Pedestrian Trajectory Prediction Method based on Decomposed Multimodal Modelling of Human Dynamics

Abstract Pedestrian trajectory prediction has wide applications in various engineering disciplines and in autonomous driving. It can effectively enhance traffic safety, improve traffic flow, optimize urban planning, and enhance the safety of autonomous driving. However, current methods can only perform unimodal predictions and cannot consider the diversity and uncertainty of pedestrian behavior. While some generative models can generate diverse prediction results, they cannot guarantee coverage of key modes and have limited control over the attributes of predicted trajectories. To address these issues, we propose a pedestrian trajectory prediction method based on Decomposed Multimodal Modeling of human dynamics, DMMNet, considering both the uncertainty of decision variables and the stochastic nature of random decision. Firstly, we adopt a decomposed modeling approach to effectively model the target uncertainty and the randomness of targets and paths in pedestrian trajectory prediction. This allows us to better capture the diversity and uncertainty of pedestrian behavior. Secondly, our method can generate explicit probability maps, providing better spatial constraints and control capabilities to improve the accuracy and interpretability of prediction results. This, in turn, offers better guidance and adaptability for other intelligent systems. Finally, our method extends the prediction range, capable of predicting pedestrian trajectories over a longer time period in the future. The proposed pedestrian trajectory prediction method in this paper has clear advantages in considering multimodality, providing spatial constraints, and expanding the application scope. It can offer more accurate, interpretable, and controllable pedestrian trajectory prediction capabilities for intelligent transportation systems and other related research and applications. The proposed method achieved an improvement of 47.7% in the ADE metric and 62.6% in the FDE metric on the ETH/UCY dataset. In the SDD dataset, there was an improvement of 18.4% in the ADE metric and 35.2% in the FDE metric.